Portable medical apps challenge low-power quest

In the design of portable medical devices, especially those that are DSP-intensive, a designer faces a number of challenges. Typically, devices must be light in weight, small in form factor and high in performance. These constraints often mean that the system must operate from a lightweight lithium ion battery or similar power source. The demands for extended battery life, sufficient computational capability and a flexible man-machine interface require the use of a DSP subsystem that is high performance (in terms of both computation and signal-conversion capabilities), ultralow power and flexible. Given the use of small batteries to meet form factor requirements, available power is also limited.

The power consumption budget for the signal-processing component of a portable medical application can typically be split into the signal conversion/conditioning and the signal processing. The power consumed by the signal conversion/conditioning is largely determined by the signal-to-noise ratio (SNR) and bandwidth required (that is, the sampling rate).

While there is certainly room for innovation in this area, generally the more current supplied to a signal-conversion front end and the larger the CMOS (or bipolar) devices used, the better the SNR. Conversion topologies that are more digital in nature, such as sigma-delta, benefit from moving to smaller semiconductor geometries. However, leakage currents and increased 1/f noise on smaller geometries also mean that more power may be needed in the signal-conditioning portions of the design to realize a desired level of performance.

The greatest reduction in power can be achieved by reducing the operating voltage. However, for a given CMOS geometry and a desired operating frequency, the operating voltage is largely specified by the threshold voltages of the selected CMOS geometry. Reducing the threshold voltages allows for reduced-voltage operation at a given clock frequency. However, subthreshold leakage currents will increase and may become a significant component of the power consumption. The reduction of leakage currents is critical in applications where a portable medical device spends long periods in standby mode.

Divide and conquer

One approach to maximizing power efficiency is to use a methodology that partitions the signal-processing algorithm into a number of blocks that can operate in parallel at a reduced clock frequency. This provides two benefits: First, blocks that are not needed can have their power cycled off. And second, the reduced clock frequency permits the use of reduced operating voltages. This is a "divide and conquer" approach that is well-known to signal-processing engineers, the best example being the ubiquitous fast Fourier transform.

An alternative divide-and-conquer approach is to partition a signal-processing algorithm block into two sets: regular, vectorized elements vs. combinatoric, irregular elements. This typically results in multirate partitioning where there is a higher-rate signal-processing path and a reduced-rate control (or side) chain. Properly applied, this approach can be used equally well in applications where the entire design is "hard-coded" in hardware (an application-specific integrated circuit, or ASIC) or in more-flexible approaches like application-specific signal processors (ASSP).

When this design approach is realized as a general-purpose digital signal processor in combination with a reconfigurable vector processor, we call it a reconfigurable application-specific signal processor. ON Semiconductor pioneered this RASSP approach for audiology applications and found that it offered the flexibility of an ASSP along with the low power consumption of an ASIC.

It is interesting to contrast this approach with the emerging dual-core (and now quad-core) homogenous approaches in general-purpose computing. Both offer reduced power consumption because they permit reduced operating frequencies and, therefore, can operate at reduced voltages. However, the dual-core heterogeneous approach offers further reductions in power, because the specialized processing needed for signal-processing applications is implemented in hardware.

The use of this approach in real-world applications has proven successful in audiology, portable audio and some emerging medical applications. A reconfigurable processor focuses on the vectorized "number crunching" and a general-purpose, dual-MAC digital signal processor handles the side-chain processing, communications and man-machine interface elements of the design. Based on this scheme, small, ultralow-power DSP applications are now emerging in a wide range of portable medical applications.

Programming challenges

Programming dual-core heterogeneous systems poses some challenges. However, with a solid tool set and proper training, most DSP engineers rise to the occasion. The use of high-level tools for programming heterogeneous dual-core systems is still an area of research, with new ideas emerging that will certainly help to streamline the process.

For the vast majority of signal-processing algorithms implemented on dual-core, heterogeneous devices, a block-based approach is used for the overall signal-processing architecture. This results in a natural system heartbeat, or "tick," which is the block rate. Partitioning a signal-processing algorithm on these systems can be challenging, because it must be partitioned across processors (the general-purpose processor vs. a reconfigurable processor) and also across time (that is, over ticks). Partitioning over ticks is used to minimize the update rate of all side-chain parameters so as to reduce power consumption to an absolute minimum.

Partitioning across processors can be simplified by the use of a Matlab-like interface to the reconfigurable vector processor and by the use of "function chains." A function chain assembles a number of vector operations into one overall function that is called from the general-purpose processor. The function chain executes on the reconfigurable vector processor in parallel with the general-purpose processor and issues an interrupt when the operation implemented is complete.

This parallel processing, in combination with the use of efficient vector processing and multirate processing, delivers the ultralow power consumption needed in portable medical applications. In short, ultralow power for portable medical applications requires a combination of the right hardware approach and an efficient, well-coded signal-processing algorithm.

Pulse oximetry example

This DSP design and programming approach has been applied to the problem of low-power, portable pulse oximetry systems. In a typical fingertip pulse oximeter, red and infrared light are alternately transmitted through the patient's finger and a photodiode is used to detect the amount of light absorbed by the finger. The ratio of absorbed red and infrared light can be used to determine the patient's blood oxygenation level (S_pO₂).

Extracting this vital sign from the received signal is a challenging signal-processing task. It's typically handled with large, off-the-shelf, general-purpose DSPs, which do not offer low power consumption needed for portable applications. Using ON Semiconductor's heterogeneous dual-core approach, the high-sample-rate signal-processing operations, such as demodulation, digital filtering, decimation and frequency domain analysis, are efficiently mapped to the reconfigurable vector processor. This offloads the general-purpose DSP to handle the complicated signal-analysis and decision logic required to determine the S_pO₂ reading. It also frees the general-purpose DSP for other system tasks, such as controlling the analog front end and operating an LCD.

In a prototype implementation of an _pO₂ system, the DSP subsystem, including signal conversion, requires only 2.5 milliamps at 1.8 volts when actively reading the S_pO₂ level.

About the authors

Todd Schneider is vice president for diagnostics, therapy and monitoring. Dave Hermann is a senior member of the technical staff for applied DSP development, in the Medical Division at ON Semiconductor.